{
 "measurement": "Whether a CLAUDE.md costs the same in Chinese as in English, for the same instructions",
 "claude_code_version": "2.1.247 (Claude Code)",
 "preregistration": "/claudemd-language-preregistration.md",
 "why": "A companion corpus measured skill descriptions at 2.2x per CHARACTER for Chinese while OVERLAPPING per BYTE. That says the per-character gap is largely a unit artefact. This tests the prediction that falls out of it, in a different place: this site prices CLAUDE.md per BYTE, so a Chinese file should cost about what an English one of the same byte size costs.",
 "method": {
  "population": "one CLAUDE.md at the root of an otherwise empty project, one turn, against an empty floor measured in the same round",
  "estimator": "guarded mode from the shared estimator, declared before the run",
  "fixture": "ONE TRANSLATION PAIR, so meaning is held constant. The bytes happen to match within 1.4% while the characters differ by 2.8x, which is what lets a single pair separate the two units.",
  "rounds": 16,
  "controls": [
   "paired floor per round",
   "ONE constant working-directory path, created and removed per run, because the startup floor moves with the path",
   "arrival control: reply exactly OK, published per run",
   "fixture byte size asserted before each run; floor arm asserts no CLAUDE.md exists",
   "--setting-sources project asserted onto the command line",
   "result_guard refuses an unusable result rather than recording a zero"
  ]
 },
 "results": {
  "headline": {
   "english_tokens": 604,
   "chinese_tokens": 678,
   "ratio_chinese_over_english": 1.123,
   "english_bytes": 1558,
   "chinese_bytes": 1536,
   "plain_statement": "The same instructions cost 678 tokens in Chinese against 604 in English, on 1536 against 1558 bytes. That is 1.12x, not the 2 to 3x a per-character reading suggests and not the 0.36x a character-count reading suggests.",
   "tokens_per_byte": {
    "en": 0.3877,
    "zh": 0.4414,
    "ratio": 1.139
   },
   "tokens_per_char": {
    "en": 0.3877,
    "zh": 1.2194,
    "ratio": 3.145
   }
  },
  "arms": {
   "en": {
    "tokens": 604,
    "rounds": 16,
    "mode_seen_in_rounds": 10,
    "min": -491,
    "max": 802,
    "spread": 1293,
    "distribution": {
     "-491": 1,
     "-95": 1,
     "406": 2,
     "604": 10,
     "802": 2
    },
    "fixture_bytes": 1558,
    "fixture_chars": 1558,
    "tokens_per_byte": 0.3877,
    "tokens_per_char": 0.3877,
    "reply_exact_ok_all_rounds": true
   },
   "zh": {
    "tokens": 678,
    "rounds": 16,
    "mode_seen_in_rounds": 12,
    "min": -219,
    "max": 876,
    "spread": 1095,
    "distribution": {
     "-219": 1,
     "480": 2,
     "678": 12,
     "876": 1
    },
    "fixture_bytes": 1536,
    "fixture_chars": 556,
    "tokens_per_byte": 0.4414,
    "tokens_per_char": 1.2194,
    "reply_exact_ok_all_rounds": true
   }
  },
  "predictions_scored": {
   "P1_site_claude_md_rate": {
    "statement": "106 tokens + 0.26 per byte, scored within +/-15%",
    "en": {
     "predicted": 511,
     "measured": 604,
     "verdict": "WRONG"
    },
    "zh": {
     "predicted": 505,
     "measured": 678,
     "verdict": "WRONG"
    },
    "NOT_A_REFUTATION": "The rate is published as fitted from 2KB to 28KB. Both fixtures here are about 1.5KB, BELOW that range, so this is the rate applied outside its declared validity and the misses are not evidence against it. Recorded because the prediction was pre-registered and dropping it would be selective reporting. It is the SECOND file on which this rate has come in low, after the markdown-structured one measured on 2.1.241, and both are under-predictions."
   },
   "P2_per_byte_rates_from_the_skill_corpus": {
    "statement": "English 0.363-0.482 and Chinese 0.459-0.513 tokens per byte, measured on skill descriptions the same day",
    "en": {
     "predicted_range": [
      566,
      751
     ],
     "measured": 604,
     "verdict": "CORRECT"
    },
    "zh": {
     "predicted_range": [
      705,
      788
     ],
     "measured": 678,
     "verdict": "WRONG",
     "note": "measured 678 against a floor of 705, i.e. 4% below the interval. The direction is right and the magnitude is close; a rate fitted on descriptions of a few hundred characters transferred to a 1.5KB file to within 4%."
    }
   },
   "P3_naive_per_character_reading": {
    "statement": "That Chinese is CHEAPER because it uses fewer characters: the headline this site had to make sure a reader could not walk away with.",
    "zh": {
     "predicted_range": [
      183,
      248
     ],
     "measured": 678,
     "verdict": "WRONG"
    },
    "how_wrong": 3.15
   }
  }
 },
 "known_limits": {
  "CARRY FORWARD the_difference_is_multimodal": "Carried from the 2.1.241 CLAUDE.md corpus, and it holds here. Neither arm is a constant: English took its mode on 10 of 16 rounds and Chinese on 12 of 16, and both wander. The excursions are QUANTISED in 198-token steps around the mode (English 406 and 802 against 604; Chinese 480 and 876 against 678), which is this machine's known 198-token intermittent component. Each arm also had one large negative round. The mode is the clean reading and the ranges OVERLAP, which is why the 74-token difference is reported as a modal difference and NOT as separated arms.",
  "CARRY FORWARD one_machine_one_model": "Opus on Claude Code 2.1.247, project scope only, one machine.",
  "CARRY FORWARD trivial_turns": "The prompt is a two-word reply, so what is isolated is the file's marginal contribution, not a session's cost.",
  "RETIRED four_turns_only": "That limit constrained a per-turn billing claim. No per-turn or billing claim is made here; this is a single-turn context measurement, so the limit does not apply rather than being carried.",
  "THE ARMS DO NOT SEPARATE, ONLY THE MODES DIFFER": "English spans -491 to 802 and Chinese -219 to 876. Those ranges overlap almost entirely. This corpus does NOT claim separated arms and no reader should take 1.12x as a resolved effect at the level of a single run. What IS supported is that the modal cost differs by 74 tokens across 16 self-paired rounds, and that every prediction of a LARGE difference in either direction is refuted.",
  "one_translation_pair_by_one_author": "One pair of files, written for this test by the same author as the rest of this site's fixtures, and NOT checked by a native speaker. A clumsy translation could tokenise worse than natural Chinese. Nothing here generalises to any other language, to a longer file, or to a file of code or tables.",
  "below_the_fitted_range": "Both fixtures are about 1.5KB and this site's CLAUDE.md rate is published as fitted from 2KB to 28KB. P1 is therefore scored as pre-registered but is explicitly NOT a refutation of that rate.",
  "no_cause_asserted": "Why Chinese costs about 14% more per byte is not measured. Byte-level fallback on CJK codepoints is a candidate and is not tested."
 },
 "runs": [
  {
   "arm": "floor",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23662,
   "delta": 480,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 22691,
   "delta": -491,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 22963,
   "delta": -219,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 22984,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 802,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23662,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 0,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23588,
   "delta": 406,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 1,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 2,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23662,
   "delta": 480,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 22984,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 802,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 3,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23662,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 4,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 4,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 4,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 5,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 5,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 5,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 6,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 6,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23588,
   "delta": 406,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 6,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 7,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 22984,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 7,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 22889,
   "delta": -95,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 7,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 876,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 8,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 8,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 8,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 9,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 9,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 9,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 10,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 10,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 10,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "floor",
   "round": 11,
   "status": "ok",
   "fixture_bytes": 0,
   "context": 23182,
   "delta": 0,
   "reply_exact_ok": true
  },
  {
   "arm": "en",
   "round": 11,
   "status": "ok",
   "fixture_bytes": 1558,
   "context": 23786,
   "delta": 604,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  },
  {
   "arm": "zh",
   "round": 11,
   "status": "ok",
   "fixture_bytes": 1536,
   "context": 23860,
   "delta": 678,
   "reply_exact_ok": true,
   "refused_because": null,
   "reply": "OK"
  }
 ]
}